Update README.md
Browse files
README.md
CHANGED
@@ -76,54 +76,31 @@ I created this model to make the hands as beautiful as possible.<br>
|
|
76 |
ぜひお試しください (Please try it!) <br>
|
77 |
<br>
|
78 |
<br>
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/654314c7c9483f8861a8e296/4ke5Y3D61gY6JlwqY_xxo.jpeg)
|
83 |
-
|
84 |
-
さらに早い生成のテストです<br>
|
85 |
-
|
86 |
-
[設定] CFG:4, STEP:6, HiresFix-Denoize:0.5(GeForce users) <br>
|
87 |
-
上の設定で試してみてください(参考プロンプトはファイルの PNG 情報からどうぞ)<br>
|
88 |
-
Here's a test of even faster generation.<br>
|
89 |
-
Try the above settings.(For reference prompts, please go to the PNG info in the file)<br>
|
90 |
-
|
91 |
-
[説明] このテストは LCM LORA を組み込んだモデルです[LCM(LORA) 0.5]<br>
|
92 |
-
その他、epi_Noize-(LORA) などをマイナス適用し生成時のノイズを減じてます<br>
|
93 |
-
※FLAT(LORA) をネガティブに入れると精細化しキレイな画像を得られます<br>
|
94 |
-
※西洋系モデルを学習元にしたノイズ LORA をマイナス適用すると西洋系の顔立ちを薄められました<br>
|
95 |
-
※東洋系モデルを学習元にしたノイズ LORA をプラス適用すると東洋系の顔立ちを濃くできました<br>
|
96 |
-
※FreeU によるチャンネル分けもマージ時の層別適用に効果的と思います<br>
|
97 |
-
皆さんの「マージ」の参考になれば幸いです<br>
|
98 |
-
[Description] This test is a model incorporating LCM LORA [LCM(LORA) 0.5].<br>
|
99 |
-
Other, epi_Noize-(LORA), etc. are applied to reduce noise during generation.[epi_Noize-(LORA) -1.0]<br>
|
100 |
-
※Negative prompt of FLAT(LORA) results in a more refined and beautiful image.<br>
|
101 |
-
※Negative noise LORA using Western models as training sources can be applied to reduce Western facial features.<br>
|
102 |
-
※Noise LORA using an oriental model as a training source can be applied to enhance oriental facial features.<br>
|
103 |
-
※Channel separation by FREEU is Effective stratified application during merging.<br>
|
104 |
-
I hope this will be helpful for your "merging"!<br>
|
105 |
<br>
|
106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
<br>
|
108 |
-
[
|
109 |
-
LCM (LORA) 全層適用の高速化は最大 62% になりましたが OUT 層への影響で消えないノイズも残りました<br>
|
110 |
-
そこで MID 層のみ LCM (LORA) 適用します、高速化は 25% に留まりますがノイズはキレイに消えます<br>
|
111 |
-
結果的にこれが最適解かなと思います ※LCM 0.5 適用の場合です強度を上げればさらに高速化するはずです<br>
|
112 |
-
(<<NON LCM STEP:16<<rev 020 LCM ALL 層 STEP:6>>LCM MID Layer STEP:12>>)<br>
|
113 |
-
[LCM Considerations] What we learned from the rev020 test<br>
|
114 |
-
LCM (LORA) applied to all layers resulted in a 62% speedup, but some noise remained due to the effect on the OUT layer.<br>
|
115 |
-
Therefore, LCM (LORA) was applied only to the MID layer. The speedup was 25%, but the noise disappeared cleanly.<br>
|
116 |
-
As a result, I think this is the best solution.<br>
|
117 |
-
※This is the case when LCM 0.5 is applied. If the intensity is increased, the speed should be further improved.<br>
|
118 |
-
(<<NON LCM STEP:16<<rev 020 LCM ALL Layer STEP:6>>LCM MID Layer STEP:12>>)<br>
|
119 |
<br>
|
120 |
-
[おすすめ]
|
121 |
-
LCM(LORA) をモデルにマージする場合 bM1280ch (FreeU 相当) にすると STEP
|
122 |
-
今回のテスト版では STEP:10 以上の場合に破綻のない画像を高頻度で得られました<br>
|
123 |
※bM1280ch = 0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0<br>
|
124 |
[Recommendation] Efficient LCM usage<br>
|
125 |
When merging LCM(LORA) into a model, set bM1280ch (equivalent to FreeU) to give more leeway for STEP setting.<br>
|
126 |
-
In this test version, we obtained images with no collapses at a high frequency when STEP:10 or higher was used.<br>
|
127 |
※bM1280ch = 0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0<br>
|
128 |
<br>
|
129 |
<br>
|
|
|
76 |
ぜひお試しください (Please try it!) <br>
|
77 |
<br>
|
78 |
<br>
|
79 |
+
|修正| マージモデル破損と修正<br>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
<br>
|
81 |
+
以前のモデルデータに「Unet 破損」を見つけたため修正しました<br>
|
82 |
+
rev01A、rev010A、TEST-rev020A、それぞれ非公開とします(ComfyUI 等で生成不能)<br>
|
83 |
+
<br>
|
84 |
+
※上の3モデルについては「マージ不適格」であることを連絡します<br>
|
85 |
+
<br>
|
86 |
+
rev010A を修正し、新たに、rev011A、rev011B-LCM、として公開します<br>
|
87 |
+
<br>
|
88 |
+
|Fixes| Merge Model Corruption and Fixes<br>
|
89 |
+
Correction of "Unet corruption" found in previous model data.<br>
|
90 |
+
rev01A, rev010A, and TEST-rev020A are now private (could not be generated by ComfyUI, etc.).<br>
|
91 |
+
※I will inform you that the above 3 models are "merge ineligible".<br>
|
92 |
+
rev010A is modified, rev011A and rev011B-LCM1-5, respectively.<br>
|
93 |
+
<br>
|
94 |
+
![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/654314c7c9483f8861a8e296/7wI8xJM4izX3R7dZCNS9c.jpeg)
|
95 |
+
![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/654314c7c9483f8861a8e296/yQT90GSxTeKXyc8LuMHYW.jpeg)
|
96 |
<br>
|
97 |
+
[Easy settings] //rev011A// STEP:16~ CFG:5 //rev011B-LCM// STEP:12~ CFG:5 あたりがオススメです<br>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
<br>
|
99 |
+
[おすすめ] 私的 LCM 活用<br>
|
100 |
+
LCM(LORA) をモデルにマージする場合 bM1280ch (FreeU 相当) にすると STEP 設定に余裕が生まれました<br>
|
|
|
101 |
※bM1280ch = 0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0<br>
|
102 |
[Recommendation] Efficient LCM usage<br>
|
103 |
When merging LCM(LORA) into a model, set bM1280ch (equivalent to FreeU) to give more leeway for STEP setting.<br>
|
|
|
104 |
※bM1280ch = 0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0<br>
|
105 |
<br>
|
106 |
<br>
|